基于句法和语义的话题细粒度情感分析的研究

发布时间：2018-08-30 17:05

【摘要】：海量互联网信息中蕴含着巨大社会价值和商业价值。传统的情感分析技术分析粒度大、准确性不高,并不能满足日益丰富的互联网用户需求。因此,如何采用一种细粒度情感分析技术使计算机能自动地对网络信息进行处理分析,给决策者提供更加详尽的结论性信息,帮助企业了解用户的消费习惯,帮助用户全面了解产品信息,逐渐成为了发展趋势。目前,针对细粒度情感分析的研究对句法和语义信息的挖掘还不够深入,因此,本文综合情感关键句抽取、评价对象识别、话题相关的情感倾向性分析的研究基础,深度挖掘句法和语义层面的知识,针对互联网语料,对细粒度情感分析进行了深入的研究。主要研究内容和创新点包括:1.针对情感关键句抽取工作,提出基于PMI的情感词典扩展算法、主题模型与词图模型相结合的关键词抽取算法和依存模板提取算法,来分别获取情感词、关键词和句法依存特征,融合词汇语义和句法依存信息完成情感关键句的抽取工作。2.针对评价对象识别工作,提出一种融合词性模板、依存结构分析、语义角色标注和短语结构分析的领域词典构建方法PDSP,并采用Word Embedding方法对该词典进行扩展之后,嵌入到序列标注模型条件随机场CRF中进行评价对象识别。3.在话题相关的情感倾向性分析中,首先提出了一种融合局部和全局信息的LTIGT算法进行关键词特征提取;然后基于Word Embedding和依存关系提取话题相关情感词,并对其进行K-means聚类,获取情感词特征;最后将以上两种特征,与传统特征一起加入SVM进行话题相关的情感倾向性判定,获得每个话题所对应的情感倾向,完成细粒度的情感分析。
[Abstract]:Mass Internet information contains huge social value and commercial value. The traditional emotion analysis technology has large granularity and low accuracy, so it can not meet the increasing demand of Internet users. Therefore, how to adopt a fine-grained emotion analysis technology to enable the computer to process and analyze the network information automatically, to provide decision makers with more detailed and conclusive information, and to help enterprises understand the consumer habits of users. Helping users to understand product information in an all-round way has gradually become a trend of development. At present, the research on fine-grained emotional analysis is not deep enough for the mining of syntactic and semantic information. Therefore, this paper synthesizes the research foundation of emotional key sentence extraction, evaluation object recognition, topic related emotional orientation analysis. Deep mining of syntactic and semantic knowledge and research on fine-grained affective analysis are carried out for the Internet corpus. The main research contents and innovations include: 1. Aiming at the extraction of emotional key sentences, this paper proposes an extension algorithm of emotion dictionary based on PMI, a keyword extraction algorithm combined with topic model and word graph model, and an algorithm of extracting dependency template to obtain emotional words, keywords and syntactic dependency features, respectively. Combining lexical semantic and syntactic dependency information to complete the extraction of affective key sentences. 2. This paper presents a domain dictionary construction method, PDSP, which combines part of speech template, dependency structure analysis, semantic role tagging and phrase structure analysis, and extends the dictionary by Word Embedding method. The evaluation object identification is carried out in conditional Random Field CRF embedded in sequence tagging model. 3. In the topic related affective orientation analysis, a LTIGT algorithm combining local and global information is proposed for keyword feature extraction, and then based on Word Embedding and dependency relation, the topic related affective words are extracted and clustered by K-means. Finally, the above two features are added to the SVM to determine the affective tendency of each topic, and the emotional tendency of each topic is obtained. Finally, the fine-grained emotional analysis is completed.
【学位授予单位】：北京理工大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.1

【相似文献】